feat(agent): false-stop detection — nudge past premature stop after tool calls

[yingliang-zhang]

Problem

When an LLM produces finish_reason=stop with text indicating intent to continue (e.g., a colon-preamble that lost its tool_calls, or a narrated continuation), the conversation silently ends without executing the intended tool call (#42503).

This is observed across model families — K3, GLM, DeepSeek, Qwen, and others. The model writes a preamble like 接下来读取文件： or 继续把剩余页面读完。 with finish_reason=stop and no tool_calls, right after a tool round completed.

Fix

Add a bounded false-stop detection mechanism that nudges the model to issue the actual tool call:

• Signature A — lost tool-call preamble: short text (<120 chars) ending with a colon (: or ：), produced after a tool round. Likely the start of a tool-call narration that lost its tool_calls.
• Signature B — narrated continuation: text (≤200 chars) ending with CJK sentence punctuation, containing at least one intent marker, and NOT containing done/wait/question markers, produced after a tool round.

Gated by agent.false_stop_detection: "auto" (on for interactive surfaces — CLI, TUI, desktop, codex, local — off for messaging platforms). 2-attempt budget, resets on tool round and genuine completion. Synthetic nudge is stripped from durable transcript and trajectory via _false_stop_synthetic ephemeral flag.

Complementary to existing features

• intent_ack_continuation (already upstream): fires at turn START (no prior tool calls) when model "announces an action but calls no tool." False-stop detection is the POST-tool-round version — it fires after a tool round completed (the _was_in_tool_round gate checks last 8 messages for assistant with tool_calls).
• _dropped_toolcall_nudge (already upstream): handles finish_reason=tool_calls with empty tool_calls array. Signature A handles finish_reason=stop with a text preamble — adjacent but distinct.
• fix(agent): nudge past progress-placeholder responses after tool call… #57610 (OPEN, "nudge past progress-placeholder responses after tool calls"): addresses a similar problem space. This PR's Signature A spec + CJK corpus was offered as a complementary contribution in a comment on fix(agent): nudge past progress-placeholder responses after tool call… #57610.

Implementation

Extracted to a dedicated agent/false_stop.py module following the agent/verification_stop.py (273 lines) and agent/kanban_stop.py (108 lines) pattern:

• build_false_stop_nudge(content, messages, attempts) — pure function returning nudge text or None
• false_stop_detection_enabled(config) — config + env gate, mirrors verify_on_stop_enabled()
• Module-level constants for marker tuples (previously rebuilt per iteration)

Marker lists are production-hardened from real deployments: 24 CJK + 13 English intent markers, with done/wait/question exclusion guards. Signature B's ending set is CJK-only (。！？;…) to avoid over-triggering on English completions ending in ..

Test plan

• scripts/run_tests.sh tests/agent/test_false_stop.py — all pass (35 tests across 5 categories: gating, Signature A, Signature B, budget/reset, transcript hygiene)
• No regression on tests/agent/test_verification_stop.py, tests/agent/test_verification_stop_caching.py, tests/run_agent/test_verification_continuation_budget.py
• py_compile on all changed files
• No circular imports